Augmenting deep neural networks with symbolic knowledge: Towards trustworthy and interpretable AI for education
Project Overview
The document explores the integration of symbolic knowledge within artificial neural networks (ANNs) to improve their effectiveness in education, addressing challenges such as educational knowledge incorporation, bias management, and system interpretability. It introduces a neural-symbolic AI approach (NSAI), which incorporates educational insights into deep neural networks, thereby enhancing the modeling of learners' computational thinking. This NSAI methodology demonstrates superior performance compared to traditional ANN methods, particularly in terms of generalizability and interpretability. The findings indicate that such an approach not only enhances the educational applicability of AI but also fosters the development of trustworthy AI applications in educational settings, ultimately aiming to create more effective and reliable learning tools.
Key Applications
Neural-Symbolic AI (NSAI) approach
Context: Educational game AutoThinking aimed at improving computational thinking skills of learners.
Implementation: Incorporates educational knowledge into deep neural networks during training to model learners' computational thinking.
Outcomes: NSAI shows better generalizability and interpretability compared to traditional deep neural networks.
Challenges: Limited application of neural-symbolic AI in education, reliance on spurious correlations in traditional ANNs.
Implementation Barriers
Technical
Difficulty in incorporating symbolic educational knowledge into ANNs.
Proposed Solutions: Utilize neural-symbolic AI frameworks to inject and extract educational knowledge.
Bias and Fairness
Deep neural networks may learn spurious correlations, leading to biases in predictions.
Proposed Solutions: Implement mechanisms to control biases, such as integrating explicit educational knowledge.
Interpretability
Lack of interpretability in traditional ANNs hampers trust and understanding among educators and students.
Proposed Solutions: Extract rules from trained networks to enhance interpretability and reasoning.
Project Team
Danial Hooshyar
Researcher
Roger Azevedo
Researcher
Yeongwook Yang
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Danial Hooshyar, Roger Azevedo, Yeongwook Yang
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai